Skill: quack

Explicit route control 🦆. Alias-first auto-route, else one targeted disambiguation question.

Purpose

Provide explicit user-controlled routing for workflow-like requests while preserving safety and approval constraints.

Philosophy Guardrails (skill-local)

Inherit shared guardrails from references/GUARDRAILS.md.

• ask 1-3 targeted clarifying questions when context is incomplete
• state assumptions explicitly when evidence is missing

Activation

Use only when user explicitly invokes quack; do not auto-activate from inferred intent.

Method

1. Verify explicit quack invocation.
2. Normalize intent: strip separator after quack (:, -, —), strip outer matching quotes, strip trailing punctuation (?!.;:).
3. Bare quack (heartbeat path):

- If normalized intent is empty or whitespace-only: - Read assets/heartbeat.md and select one random heartbeat line from the list - Read and emit full assets/quick-help.md content verbatim (no summarization) - Emit closing prompt: What would you like to route? - Stop (do not generate ad-hoc quips, do not enter disambiguation)

1. Parse optional override: use <subagent> or with <subagent> or via <subagent>.
2. Validate override (if present) against platform subagent list; if invalid, ask one correction question: Need one detail: unknown subagent "<x>". Use duckling or general? and stop.
3. Determine effective subagent: if override parsed in step 4, use that; otherwise default to duckling.
4. Determine route execution policy:

- inline-default: duck-design, duck-teach, duck-debug, duck-debt - delegated-default: duck-patch, duck-refactor, duck-risk, duck-review, duck-triage, duck-simplify - user override forces delegation regardless of default policy

1. If explicit full skill name in input, resolve that skill and execute route (see step 11).
2. Otherwise, load assets/route-aliases.json and attempt case-insensitive match.
3. Apply keyword-based precedence (when multiple skills match):

- Scan intent for precedence keywords: - Risk signals (rollback, breaking, compatibility, compat, failure, impact) -> prioritize duck-risk - Complexity signals (overengineered, complexity, simpler, too many, bloated) -> prioritize duck-simplify - Learning signals (why, how, what does, explain, teach me) -> prioritize duck-teach over duck-debug - Test signals (test, coverage, should I test, before PR) -> prioritize duck-triage - Design signals (choose, tradeoff, approach, option) -> prioritize duck-design - Reorder matched skills with precedence-matched skill first - Still present alternatives (precedence aids ranking, doesn't eliminate choice)

1. If multiple aliases match after precedence, apply tie-break: exact match > longest alias > ask one disambiguation question.
2. Route execution (explicit skill name or alias match):

- Load role instructions from assets/subagent-runbook.md for resolved skill - If inline-default policy AND no user override: - execute skill inline first - if execution fails, ask one corrective question and stop - emit Routing: <skill>. only after inline execution step completes - Otherwise (delegated-default OR user override): - delegate to subagent using harness-native dispatch tool with subagent=<effectivesubagent> (determined in step 6, defaults to duckling) first - pass skillname=<resolved_skill> parameter - if dispatch fails, ask one corrective question and stop - emit Routing: <skill> via <subagent>. if user override supplied, else Routing: <skill>. only after dispatch succeeds - Do not stop at routing text alone when execution path is available.

1. Alias miss (disambiguation):

- Detect intent fragment and ask one targeted question: - debug-ish (error/fail/trace/stack/broken): Need one detail: is this debug, trace, or review? - rollout/risk (rollout/migration/compat/rollback): Need one detail: is this risk review or design tradeoff? - code-change (fix/change/refactor/clean up): Need one detail: do you want review, patch, refactor, or simplify? - tech-debt (todo/defer/fixme/debt): Need one detail: is this debt audit or simplify? - unknown: Need one detail: which route fits—review, debug, design, teach, triage, risk, simplify, patch, refactor, or debt? - Wait for user clarification, then retry alias resolution.

Boundaries

• Preserve user decision ownership.
• If a change would weaken trust-boundary validation, security controls, data-loss prevention, accessibility requirements, or explicit user requirements, refuse it and offer only a safe alternative preserving the constraint.
• No edits/mutating commands/subagent delegation that changes workspace state without explicit bounded approval.